Finetuning LLava NeXT 34B

[akshatbhandari15]

I am trying to finetune LLava NeXT 34B
Hardware 4x NVIDIA A100 80GB

Issue: The training process is OOMing. Am I missing something in setting up distributed training?

I am using the following training config:

export OMP_NUM_THREADS=8
export NCCL_IB_DISABLE=0
export NCCL_IB_GID_INDEX=3
export NCCL_SOCKET_IFNAME=eno1
export NCCL_DEBUG=INFO

LLM_VERSION="Qwen/Qwen1.5-32B-Chat"
# for 7b model we recommend bs=1, accum=2, 16 nodes, 128 gpus, lr=1e-5, warmup=0.03
# for 72b model we recommend bs=1, accum=1, 32 nodes, 256 gpus, lr=1e-5, warmup=0.03
LLM_VERSION_CLEAN="${LLM_VERSION//\//_}"
#VISION_MODEL_VERSION="/home/ashutosh/.cache/huggingface/hub/models--google--siglip-so400m-patch14-384/snapshots/9fdffc58afc957d1a03a25b10dba0329ab15c2a3"
VISION_MODEL_VERSION="google/siglip-so400m-patch14-384"
VISION_MODEL_VERSION_CLEAN="${VISION_MODEL_VERSION//\//_}"

############### Pretrain ################

BASE_RUN_NAME="llavanext-google_siglip-so400m-patch14-384-Qwen_Qwen2-34B-Instruct-mlp2x_gelu-pretrain_blip558k_plain"
echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"

############### Finetune ################

# Stage 2
PROMPT_VERSION="qwen_1_5"
RUN_NAME="llava-next-34b-qwen-aux-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-si_stage_am9-3-epoch-full-tune" 

PREV_STAGE_CHECKPOINT="lmms-lab/llava-next-qwen-32b"

echo "PREV_STAGE_CHECKPOINT: ${PREV_STAGE_CHECKPOINT}"
echo "MID_RUN_NAME: ${RUN_NAME}"

RANK=${RANK:-0}
ADDR=${ADDR:-"127.0.0.1"}
PORT=${PORT:-"29502"}
NNODES=${NNODES:-4}
NUM_GPUS=${NUM_GPUS:-4}

ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NNODES}" --node_rank="${RANK}" --master_addr="${ADDR}" --master_port="${PORT}" \
    llava/train/train_mem.py \
    --deepspeed scripts/zero3.json \
    --model_name_or_path $PREV_STAGE_CHECKPOINT \
    --version $PROMPT_VERSION \
    --data_path="/path/to/data/llava_instruct/llava1_6mix.json" \
    --image_folder /path/to/data/llava_data \
    --mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" \
    --mm_vision_tower_lr=2e-6 \
    --vision_tower ${VISION_MODEL_VERSION} \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --group_by_modality_length True \
    --image_aspect_ratio anyres_max_9 \
    --image_grid_pinpoints  "(1x1),...,(6x6)" \
    --mm_patch_merge_type spatial_unpad \
    --bf16 True \
    --run_name $RUN_NAME \
    --output_dir /home/anamika/LLaVA-NeXT/output/$RUN_NAME \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 3000 \
    --save_total_limit 5 \
    --learning_rate 1e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 32768 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --torch_compile True \
    --torch_compile_backend "inductor" \
    --dataloader_drop_last True \
    --frames_upbound 32 \
    --quant_type "nf4" \
    --lora_enable True \
    --lora_r 64 \
    --lora_alpha 16
exit 0;